ultrawork-graph
/ultrawork-graph
Dzieli dużą pracę na niezależne węzły wykonywane przez równoległe agenty i scala wyniki po weryfikacji każdego węzła.
- Autor
- codicovibes ↗
- Repozytorium
- devstack ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 05.11.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 1
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 05.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
Co robi i kiedy
Dzieli dużą pracę na niezależne węzły wykonywane przez równoległe agenty i scala wyniki po weryfikacji każdego węzła.
Wyzwalacze z opisu
Oryginalny opis (EN)
Runs a large job as a graph of Claude Code subagents — splits the work into independent nodes, runs them in parallel on the model that fits each node, has a separate verifier check every node, and merges the results once. Use when the prompt contains "mass ulw" or "mass ultrawork", or asks to parallelize, fan out, split across agents or run in parallel, and for migrations or refactors spanning many independent files or packages, audits of many modules, or research across many sources. Not for a single change or bug fix, which the ultrawork skill owns, and not for designing a knowledge graph, w
Jak działa
- 1. Decide the shape before spawning anything
- 2. Draw the graph
- 3. Run the nodes
- 4. Verify each node, then merge once
- 5. Report
- Release valve
Podgląd SKILL.md
--- name: ultrawork-graph description: Runs a large job as a graph of Claude Code subagents — splits the work into independent nodes, runs them in parallel on the model that fits each node, has a separate verifier check every node, and merges the results once. Use when the prompt contains "mass ulw" or "mass ultrawork", or asks to parallelize, fan out, split across agents or run in parallel, and for migrations or refactors spanning many independent files or packages, audits of many modules, or research across many sources. Not for a single change or bug fix, which the ultrawork skill owns, and not for designing a knowledge graph, which graph-engineering owns. metadata: version: 0.1.0 concept: OmO "mass ulw" (oh-my-openagent), reimplemented on Claude Code subagents; design rules from graph-engineering references/task-graphs.md --- # Ultrawork graph The shape of the work decides whether a team helps. Parallel agents win on work that splits into independent pieces and lose on sequential work, so the first job is to decide the shape honestly, and the second is to keep every node checkable. Read `../graph-engineering/references/task-graphs.md` (fake edges, the diamond, the stop rule, the human gate) before drawing the first graph in a session. ## 1. Decide the shape before spawning anything Scout first: use the contextplus tools (`get_context_tree`, `semantic_navigate`, `get_blast_radius`) and Hindsight's knowledge pages to learn what the job touches. Then apply the stop rule: - The work splits into pieces whose outputs do not feed each other → build a graph. - Most steps need the previous step's output, or every piece edits the same files → it is sequential. Run the ultrawork skill in this context instead, and say in one line that you did and why. - Fewer than three independent pieces → the coordination costs more than it saves; run the ultrawork skill. ## 2. Draw the graph Each node is one job with: a goal, the files or sources it owns, its done-evidence (the command or check that proves it, run in the node's own context), and its agent type. Draw an edge only where one node's output is another node's input; delete any edge you cannot name the payload of. Two nodes that would write the same file are one node. Record the graph in the task list (one task per node, dependencies as blockers) so the user can watch it, and write it to `.claude/ulw/<slug>/graph.md` in the repo when the job will outlive this session. Place a human gate only before a step that is expensive to undo — a schema migration, a deletion, a push, anything that leaves the machine. Everything else runs without asking. ## 3. Run the nodes | Node kind | Agent | Why | |---|---|---| | Read-only discovery, inventory, source research | `devstack:scout` (Haiku) | cheap and fast; cannot edit | | A code change | `devstack:worker` (Sonnet, own git worktree) | follows the ultrawork skill; isolation prevents write conflicts | | Checking a node's output | `devstack:verifier` (session model, cannot edit) | a fresh context that did not produce the work | Launch every node whose blockers are done in one message, so they run in parallel. Give each agent a self-contained brief: the goal, the files it owns, the files it must not touch, its done-evidence, and the report shape in §5. A subagent sees none of this conversation. ## 4. Verify each node, then merge once Send each finished node to a verifier with the node's goal, its done-evidence and the worker's report. A node is done only when the verifier ran the evidence and it passed. A node that fails verification goes back to a fresh worker with the verifier's findings, at most twice. After the second failed retry, stop retrying it, keep its worktree, and carry it into the report as unresolved. Merge in this context: bring the verified worktrees onto the branch one at a time, resolve conflicts yourself, then run the repo's own test and lint commands across the merged result. The merge is not done until those pass or the failures are named. ## 5. Report Lead with what is not done: unresolved nodes with the verifier's last finding, merge conflicts or failing checks, and gates waiting on the user. Then list each completed node with its evidence (command and result), and the worktrees left on disk. Out-of-scope findings go last as follow-ups; do not fix them inside this job. Do not write memory by hand at the end. Hindsight retains the session automatically. Store something explicitly only when the user asks. ## Release valve For a small fan-out (three or four read-only nodes, no code changes) skip the graph file and the verifiers, run scouts in parallel, and say that you compressed the process.
Wersje
W bazie nie ma innych skilli o nazwie ultrawork-graph.
Migawka bazy z 13.11.2026 · wersja paczki aa1b7954aac3 (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 100ultrawork
Uruchamia równoległe wykonywanie zadań o dużej przepustowości, gdy użytkownik prosi o równoległość lub o ulw. - 100ultrawork
Uruchamia równoległe zadania agentów, gdy trzeba wykonać kilka czynności naraz. - 100ultrawork
Dzieli wieloetapowe zadania na równoległe podzadania, kieruje je do najtańszych wystarczających modeli i weryfikuje wyniki. - 85ultrawork
Wykonuje jedną zmianę w kodzie z dowodem na każdym kroku: czyta kod, odtwarza błąd, robi najmniejszą poprawkę i sprawdza wynik poleceniami w tej sesji. - 84team-shinchan:ultrawork
Szybko wykonuje zadania przez równoległe uruchamianie agentów, gdy liczy się czas realizacji. - 100ultrawork
Uruchamia pełny potok specify i execute jednym poleceniem, łącząc oba skille w automatyczny przepływ end-to-end.